--- title: "Vision Models: Document AI, Image Generation, Screen Grounding" description: "A senior lab's read on vision models: image generation pricing, document understanding across Claude and Gemini, and open-weight OCR specialists." source: "https://www.kensink.com/models/vision/" canonical: "https://www.kensink.com/models/vision/" --- ★ Vision LLM Models 8-week engagement VISION · DOCUMENTS, IMAGES, SCREENS # Four jobs that share one word. Reading a document, generating an image, grounding a click on a screen, and segmenting an object are separate problems with separate leaders. The frontier text models now do most of the reading, specialist parsers still win on dense tables, and the cheapest image model has the widest feature set. We route per job rather than per vendor. Vision API OCR engine Eval pipelines [Start a conversation →](https://www.kensink.com/contact) [All llm models →](https://www.kensink.com/models) Cycle 8 weeks · fixed price Stack Vision + document AI Output Production code + eval suite Handoff Full source ownership \[THE SHORT VERSION\] ## Half the confusion is a team choosing the wrong family. Generation models make pictures and cannot read them. Vision-language models read and cannot draw. Document parsers extract structure and cannot reason about it. Teams routinely pick an image model for a comprehension job because both are called vision, then spend a sprint discovering the mistake. We settle which of the four jobs you actually have before anyone opens a pricing page. When it fits - Extracting structure from invoices, filings, forms, and contracts - Reading dense charts and tables nested inside PDFs - Agents that operate a real browser or desktop interface - Generating and editing product, marketing, or catalogue imagery - Segmentation and object boundaries in an image pipeline When it does not - Documents with a fixed layout that a deterministic parser already handles - Anything where a wrong extraction reaches a customer without a review step \[HOW WE BUILD IT\] ## How we build with Vision. 01 ### Scope and fit We decide where Vision earns its place in your system, and where a simpler tool wins. No resume-driven architecture. 02 ### Build on a tested foundation We integrate Vision against a foundation we trust: typed code, CI, and observability from the first commit. Boring infrastructure, modern surface. 03 ### Eval before launch An eval suite proves the build behaves before it reaches a user. We measure, then ship. 04 ### Handoff with ownership Your team gets the code, the tests, and a runbook. No lock-in to us or to a vendor framework. \[ WHAT YOU GET \] ## What the engagement leaves behind. Senior Engineers who have shipped this before 100% Source ownership at handoff Eval-first Tested before it ships 0 Framework lock-in \[MODELS + VENDORS\] ## Four jobs, four different leaders. Document understanding, image generation, screen grounding, and segmentation are separate problems that share a word. No vendor leads all four, and the frontier text models now do a lot of the reading, so the routing decision matters more here than the model does. We integrate them directly behind one abstraction and pick per job. [ Flagship OpenAI Apr 2026 ### GPT-Image-2 gpt-image-2 Image generation and editing Image out $30 / 1M Image in $8 / 1M - Token-priced rather than per-image, so cost tracks resolution and detail - Generation, edits, and inpainting, with a batch path at half rates - Image input cached at $2 per million, a quarter of the uncached rate Read the technical brief ](https://www.kensink.com/models/vision/gpt-image-2)[ Flagship Google 2026 ### Nano Banana Pro gemini-3-pro-image Image generation and editing, top tier 1K / 2K image $0.134 4K image $0.24 - Per-image pricing you can quote to a client before you build - Up to six object reference images, semantic-mask editing, 1K to 4K output - Legible stylised text, which is the failure mode that used to rule generation out Read the technical brief ](https://www.kensink.com/models/vision/gemini-3-pro-image) Current Google 2026 ### Nano Banana 2 gemini-3.1-flash-image Image generation, high volume 1K image $0.067 4K image $0.151 - Half the price of Pro at 1K, with a wider reference-image surface - Up to 10 object, 4 character, and 3 style references, plus ten aspect ratios - 512px output as well, which Pro does not offer In the routing table · brief not yet published Preview Google 2026 ### Gemini 3.1 Pro gemini-3.1-pro-preview Document understanding at volume Billing Per token, preview - Native PDF input, so no image-conversion stage in your pipeline - Leads OCR and visual question answering on the Nanonets IDP leaderboard - Handles sparse tables at 94%, the case that breaks most document pipelines In the routing table · brief not yet published Flagship OpenAI 3 Sep 2026 ### GPT-6 Astra gpt-6-astra Screen grounding and computer use Input $10 / 1M Output $50 / 1M - 92.7% on ScreenSpot-Pro against 76.9% for GPT-5.6 Sol - The model to reach for when an agent has to click a real interface - Image in, text out. It reads screens, it does not draw them In the routing table · brief not yet published Flagship Anthropic 1 Sep 2026 ### Claude Fable 5.1 claude-fable-5-1 Dense documents and charts Input $10 / 1M Output $50 / 1M - Built for dense charts, filings, and tables nested inside PDFs - Crop-and-zoom tools on charts, which is how a human reads them too - 1M context, so a long filing and its exhibits fit in one pass In the routing table · brief not yet published Current Open weights 2026 ### PaddleOCR VL 1.5 PaddleOCR-VL-1.5 Document parsing specialist Billing Self-hosted - Tops OmniDocBench at 94.37 overall, ahead of every general model on that test - Self-hostable, which settles the question when documents cannot leave your network - A parser, not a reasoner. Pair it with an LLM for anything beyond structure In the routing table · brief not yet published Current Open weights 2026 ### Qwen3-VL Qwen3-VL General vision language model Billing Self-hosted - Competitive with proprietary frontier models across multimodal benchmarks - Covers grounding, video understanding, OCR, and document comprehension - The open-weight default when you need vision inside your own boundary In the routing table · brief not yet published Current Meta 2026 ### SAM 3 sam-3 Segmentation Billing Self-hosted - Segment anything: masks and object boundaries rather than descriptions - The right tool when you need pixels back, not prose - Pairs with a VLM, which names what SAM has isolated In the routing table · brief not yet published \[METHODOLOGY · K-FRAMEWORK\] ## Integrated through the K-Framework. Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship. [Read the K-Framework](https://www.kensink.com/k-framework) 01 ### Foundations Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres. 02 ### Amplification An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo. 03 ### Judgment Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots. \[OBSERVABILITY\] ## Observability your team can read. A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket. Instrumented ### Cost per call Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters. Instrumented ### Latency p50 / p95 / p99 Real distributions, not averages. We know which routes are slow, and why. Instrumented ### Eval pass rates The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast. Instrumented ### Prompt + completion logs PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window. Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics. \[COMMON QUESTIONS\] ## Questions we get asked. Which model reads documents best? It depends on the document. Gemini 3.1 Pro leads OCR and visual question answering on the Nanonets leaderboard and takes native PDF input, Claude Fable 5.1 is built for dense charts and tables nested inside filings, and PaddleOCR VL 1.5 tops OmniDocBench at 94.37 as a dedicated parser. We build an eval set from the customer's own documents before choosing, because the differences show up on layout rather than on a leaderboard. Can an image generation model read an image? No, and this is the most common mistake in the category. GPT-Image-2 and Nano Banana generate and edit; they do not describe, extract text, or ground a click. Reading goes to text models with image input. Both are called vision, and they are entirely different families. What does image generation actually cost? It depends on how the vendor bills. Nano Banana Pro is $0.134 per 1K or 2K image and $0.24 at 4K, a number you can put in a quote. GPT-Image-2 bills output at $30 per million tokens with cost scaling by resolution and detail, so you have to measure your own prompts first. For fixed-price work that difference usually decides it. Do generated images carry a watermark? Google applies SynthID to every image its models generate, and Anthropic ships C2PA Content Credentials on media produced through its code execution tool. Confirm the position for whichever model you choose rather than assuming, because provenance requirements are tightening under the EU AI Act. Should we pin a model version? Always. Image models shift their default aesthetic between snapshots, nothing fails and no test goes red, and the assets simply stop matching the ones already approved. Pin the snapshot, keep a reference set of approved output, and diff before rolling a version forward. Share[](https://twitter.com/intent/tweet?url=https%3A%2F%2Fwww.kensink.com%2Fmodels%2Fvision%2F&text=Vision%20%C2%B7%20LLM%20Models)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.kensink.com%2Fmodels%2Fvision%2F) [View .md](https://www.kensink.com/models/vision.md) \[RELATED\] ## Worth a look next. [ LLM Models Voice & Speech Read more ](https://www.kensink.com/models/whisper-speech)[ LLM Models Claude Read more ](https://www.kensink.com/models/claude)[ Design Patterns RAG Read more ](https://www.kensink.com/llm/rag) APPLIED K-FRAMEWORK ## Bring the problem. We’ll bring the build. Senior engineers, eval suite at handoff, full source ownership. Sprint, program, or ongoing. We shape the engagement to the work. [Start a conversation →](https://www.kensink.com/contact) [Read the K-Framework](https://www.kensink.com/k-framework)